def getFeature(net, dataloader, device, flip = True):
### Calculate the features ###
featureLs = None
featureRs = None
count = 0
for det in dataloader:
for i in range(len(det)):
det[i] = det[i].to(device)
count += det[0].size(0)
# print('extracing deep features from the face pair {}...'.format(count))
with torch.no_grad():
res = [net(d).data.cpu() for d in det]
if flip:
featureL = l2_norm(res[0] + res[1])
featureR = l2_norm(res[2] + res[3])
else:
featureL = res[0]
featureR = res[2]
if featureLs is None:
featureLs = featureL
else:
featureLs = torch.cat((featureLs, featureL), 0)
if featureRs is None:
featureRs = featureR
else:
featureRs = torch.cat((featureRs, featureR), 0)
return featureLs, featureRs
def prepare_facebank(model, path='facebank', tta=True):
model.eval()
embeddings = []
names = ['']
data_path = Path(path)
for doc in data_path.iterdir():
if doc.is_file():
continue
else:
embs = []
for files in listdir_nohidden(doc):
try:
image_path = os.path.join(doc, files)
img = cv2.imread(image_path)
if img.shape != (112, 112, 3):
bboxes, landmarks = create_mtcnn_net(
img,
12,
device,
p_model_path='MTCNN/weights/pnet_Weights',
r_model_path='MTCNN/weights/rnet_Weights',
o_model_path='MTCNN/weights/onet_Weights')
img = Face_alignment(img,
default_square=True,
landmarks=landmarks)
if img == []:
continue
with torch.no_grad():
if tta:
mirror = cv2.flip(img[0], 1)
emb = model(
test_transform(img[0]).to(device).unsqueeze(0))
emb_mirror = model(
test_transform(mirror).to(device).unsqueeze(0))
embs.append(l2_norm(emb + emb_mirror))
else:
embs.append(
model(
test_transform(
img[0]).to(device).unsqueeze(0)))
except:
continue
if len(embs) == 0:
continue
embedding = torch.cat(embs).mean(0, keepdim=True)
embeddings.append(embedding)
names.append(doc.name)
embeddings = torch.cat(embeddings)
names = np.array(names)
print(names)
torch.save(embeddings, os.path.join(path, 'facebank.pth'))
np.save(os.path.join(path, 'names'), names)
return embeddings, names
t1 = time.time() - t
t = time.time()
faces = Face_alignment(image, default_square = True,landmarks = landmarks)
embs = []
test_transform = trans.Compose([
trans.ToTensor(),
trans.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
for img in faces:
if args.tta:
mirror = cv2.flip(img,1)
emb = detect_model(test_transform(img).to(device).unsqueeze(0))
emb_mirror = detect_model(test_transform(mirror).to(device).unsqueeze(0))
embs.append(l2_norm(emb + emb_mirror))
else:
embs.append(detect_model(test_transform(img).to(device).unsqueeze(0)))
source_embs = torch.cat(embs) # number of detected faces x 512
diff = source_embs.unsqueeze(-1) - targets.transpose(1, 0).unsqueeze(0) # i.e. 3 x 512 x 1 - 1 x 512 x 2 = 3 x 512 x 2
dist = torch.sum(torch.pow(diff, 2), dim=1) # number of detected faces x numer of target faces
minimum, min_idx = torch.min(dist, dim=1) # min and idx for each row
min_idx[minimum > ((args.threshold-156)/(-80))] = -1 # if no match, set idx to -1
score = minimum
results = min_idx
# convert distance to score dis(0.7,1.2) to score(100,60)
score_100 = torch.clamp(score*-80+156,0,100)
image = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
def get_frame(self): #여기다가 face detection, recognition기능을 넣으면 문제없음.
frame = self.frame #thread가 update하는 이미지를 가져 옴.
if frame is not None and self.flag is True:
try:
start_time=time.time()
input=resize_image(frame,self.scale)# input size를 줄여줌으로 speed up 가능
#print('get bboxes')
# bboxes, landmarks = create_mtcnn_net(input, self.min_face, self.device, p_model_path='MTCNN/weights/pnet_Weights',
# r_model_path='MTCNN/weights/rnet_Weights',
# o_model_path='MTCNN/weights/onet_Weights')
#print('sucess bbox')
bboxes,landmarks=self.face_detector.detect_all_net(image=input,mini_face=self.min_face)
if bboxes != []:
bboxes=bboxes/self.scale
landmarks=landmarks/self.scale
faces= Face_alignment(frame,default_square=True,landmarks=landmarks)
embs=[]
test_transform = trans.Compose([
trans.ToTensor(),
trans.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
for img in faces:
if self.tta:
mirror = cv2.flip(img,1)
emb = self.detect_model(test_transform(img).to(self.device).unsqueeze(0))
emb_mirror = self.detect_model(test_transform(mirror).to(self.device).unsqueeze(0))
embs.append(l2_norm(emb + emb_mirror))
else:
embs.append(self.detect_model(test_transform(img).to(self.device).unsqueeze(0)))
source_embs=torch.cat(embs)
diff=source_embs.unsqueeze(-1) - self.targets.transpose(1, 0).unsqueeze(0) # i.e
dist = torch.sum(torch.pow(diff, 2), dim=1) # number of detected faces x numer of target faces
minimum, min_idx = torch.min(dist, dim=1) # min and idx for each row
min_idx[minimum > ((self.threshold-156)/(-80))] = -1 # if no match, set idx to -1
score = minimum
results = min_idx
score_100 = torch.clamp(score*-80+156,0,100)
FPS=1.0/(time.time()-start_time)
cv2.putText(frame,'FPS : {:.1f}'.format(FPS),(10,15),cv2.FONT_HERSHEY_DUPLEX,0.75,(255,0,255))
for i,b in enumerate(bboxes):
b=b.astype('uint32')
cv2.rectangle(frame,(b[0],b[1]),(b[2],b[3]),(0,255,0),1)
try:
if self.names[results[i]+1]=='Unknown': #mosic func
#print('detect unknwon')
face_region=frame[b[1]:b[3],b[0]:b[2]]
face_region=cv2.blur(face_region,(30,30))
frame[b[1]:b[3],b[0]:b[2]]=face_region
except:
pass
# development version
# if self.score:
# cv2.putText(frame,self.names[results[i]+1]+' score:{:.0f}'.format(score_100[i]),(int(b[0]),int(b[1]-25)),cv2.FONT_ITALIC,1,(255,255,0))
# else:
# cv2.putText(frame,self.names[results[i]+1],(int(b[0]),int(b[1]-25)),cv2.FONT_ITALIC,1,(255,255,0))
except:
pass
_, jpeg = cv2.imencode('.jpg',frame)
return jpeg.tobytes()